Re: btree on disk

On 7/10/07, Joel Nelson <[email protected]> wrote:
> > it otherwords, I have to check if a graph contains a node n, if it does not
> > write a new node, and if it does retrieve the node. I need to be able to do
> > this very fast. I have tried object oriented dbms, although they worked
> > nicely for graph things, they were very poor at identity search.
> >
>
> Remember what a btree is good at: requiring a minimal number of disk
> accesses to do range queries, e.g. "give me all nodes with a key
> between 54-98"
>
> If you do not need range queries, and you don't need to iterate over
> the nodes in some sorted order, there is no advantage I know of to
> using a b-tree, or any other tree structure.
>
> Instead, it sounds like you already know what exact key value to look
> for (there is no range). In this case, a hash table (and therefore the
> HashFiler) may be a lot faster for you. A hash table can be configured
> so that it almost always finds (or reports that it cannot find) your
> node with 1 disk access, versus a btree which can require 3-5.

While it is true, hash table is not always a preferred data structure
because as more values are added, it will get collisions more often
and will get slower. I am not sure if it appears anywhere in Xindice
documentation, but general guidelines for selecting between HashFiler
and BTreeFiler are the following:

1. If amount of records that are going to be stored can be estimated
and won't change too much, HashFiler is a way to go. In this case
filer can be configured to allocate optimal amount of space for fast
data access.

2. If amount of records is unknown or likely to grow or hash function
does not work too well, BTreeFiler will show better performance.

This applies to any kind of data.

> Also depending on what you are trying to do with these graphs, there
> are specialized data structures that may work even better than a hash
> table. If you are willing as you said to create your own database, you
> should not be afraid to investigate using your own file format.
>
> Here are some common structures used for storing graphs:
>
> http://en.wikipedia.org/wiki/Adjacency_list
> http://en.wikipedia.org/wiki/Adjacency_matrix
>
> Hope this helps.
>